|new 0.2.2||May 4, 2021|
|0.2.1||May 4, 2021|
|0.2.0||Nov 19, 2019|
|0.1.1||May 11, 2019|
|0.1.0||May 7, 2019|
#418 in Procedural macros
2,215 downloads per month
Used in 2 crates (via gc-arena)
This repo is home to the
gc-sequence crates, which provide Rust
with garbage collected arenas and a means of safely interacting with them.
These crates are still fairly experimental and WIP, but they do work and provide genuinely safe garbage collected pointers.
gc-arena crate, along with its helper crate
safe allocation with cycle-detecting garbage collection within a closed
"arena". There are two techniques at play that make this system sound:
Garbage collected objects are traced using the
Collecttrait, which must be implemented correctly to ensure that all reachable objects are found. This trait is therefore
unsafe, but it can safely be implemented by procedural macro, and the
gc-arena-deriveprovides such a safe procedural macro.
In order for garbage collection to take place, the garbage collector must first have a list of "root" objects which are known to be reachable. In our case, the user of
gc-arenachooses a single root object for the arena, but this is not sufficient for safe garbage collection. If garbage collection were to take place when there are garbage collected pointers anywhere on the Rust stack, such pointers would also need to be considered as "root" objects to prevent memory unsafety.
gc-arenasolves this by strictly limiting where garbage collected pointers can be stored, and when they can be alive. The arena can only be accessed through a single
mutatemethod which takes a callback, and all garbage collected pointers inside this callback are branded with an invariant lifetime which is unique to that single callback call. Thus, when outside of this
mutatemethod, the rust borrow checker ensures that it is not possible for garbage collected pointers to be alive anywhere on the stack, nor is it possible for them to have been smuggled outside of the arena's root object. Since all pointers can be proven to be reachable from the single root object, safe garbage collection can take place.
In other words, the
gc-arena crate does not retrofit Rust with a globally
accessible garbage collector, rather it only allows for limited garbage
collection in isolated garbage collected arenas. All garbage collected pointers
must forever live inside only this arena, and pointers from different arenas are
prevented from being stored in the wrong arena.
gc-arena crate, while still useful, is very limiting. Since garbage
collection can only take place between calls to
mutate, arbitrarily long
mutate could prevent garbage collection for an arbitrary amount of
gc-sequence attempts to ease this problem by supplying a
futures-like API on
top of garbage collected arenas. It allows users to build up
state machines, using combinators, which are driven forward by repeatedly
Sequence::step. The resulting
Sequence state machines themselves
Collect, so the provided arena wrapper can safely execute
arbitrarily long running
Sequences, garbage collecting in-between calls to
Sequence::step. In this way, the
Sequence trait becomes a way to express
garbage collector "safe points", at which collection can safely take place.
These crates were developed primarily as a means of writing VMs for garbage collected languages in safe Rust, but there are probably many more uses than just this.
Currently, these crates are WIP and experimental, but are basically usable and safe. Some notable current limitations:
While these crates hopefully contain no pathalogical slowness, they are not highly optimized currently. The garbage collector in the
gc-arenacrate is a basic incremental mark-and-sweep collector which by itself is not terrible, but there is still a lot of unnecessary space overhead per-allocation. Additionally, there is not currently a way to allocate DSTs in a Gc pointer, instead requiring very slow double indirection for any non-
Sizedtype. Both of these problems are very solvable, but this work hasn't been done yet.
There is currently no system for object finalization. This is not terribly difficult to implement, depending on the system, but it would require picking a particular set of edge-case finalization behavior.
There are lots of missing
Sequencecombinators, and there is currently no way of conveniently constructing looping
Sequences. This is easily solvable, but not currently done.
A harder to solve limitation is that there is currently no system for multi-threaded allocation and collection. The basic lifetime and safety techniques here would still work in an arena supporting multi-threading, but these crates do not support this.
Another limitiation is that the
Collecttrait does not provide a mechanism to move objects once they are allocated, so this limits the types of collectors that could be written.
gc-sequencecrate is inconvenient to use, as it requires combinators to build up
Sequences. This could be made vastly easier with the help of Rust generators, BUT there would still be the limitation that generators would not automatically implement
Collect. Therefore, even with generator support, it would still require avoiding keeping garbage collected values kept inside the generators, limiting their usefulness. If it were ever possible to implement
Collecton generators automatically through some as-yet-undesigned rust feature, this would greatly improve the convenience and viability of the techniques in these crates in general.
The crates are currently very light on documentation and examples.
The ideas here are mostly not mine, much of the design is borrowed heavily from
the idea of using "generativity" comes from You can't spell trust without
the vast majority of the
Sequence design is taken directly from
Everything in this repository is licensed under either of:
- MIT license LICENSE-MIT or http://opensource.org/licenses/MIT
- Creative Commons CC0 1.0 Universal Public Domain Dedication LICENSE-CC0 or https://creativecommons.org/publicdomain/zero/1.0/
at your option.